CN105957028A - GPU acceleration patch-based bilateral filter method based on OpenCL - Google Patents

Abstract

The invention discloses a GPU acceleration patch-based bilateral filter method based on OpenCL. The GPU acceleration patch-based bilateral filter method is used to solve technical problems of conventional patch-based bilateral filter methods such as low filter efficiency and weak universality. The realization steps are as follows: 1, an isomerous architecture is built; 2, a host terminal is used to read an original image to a host memory, and is used to allocate host memory for a processed image; 3, a Gaussian template is generated by the host terminal; 4, cache objects are created by the host terminal; 5, a kernel function is written; 6, kernel parameters are set by the host terminal; 7, parameters required by kernel entry are set by the host terminal; 8, the kernel function enters the host terminal; 9, a device terminal is used to calculate range Gaussian template; 10, the device terminal is used for normalize weighted average calculation, and is used to write the data of the processed image to the corresponding cache object; 11, the data of the processed image is acquired. The GPU acceleration patch-based bilateral filter method is advantageous in that the filtering efficiency is high, the universality is strong, and the real-time processing of the high-resolution image is realized.

Description

GPU based on OpenCL accelerates burst bilateral filtering method

Technical field

The present invention relates to Digital Signal Processing and Heterogeneous Computing technical field, relate to a kind of burst bilateral filtering method, tool Body relates to a kind of GPU based on OpenCL and accelerates burst bilateral filtering method, can be used for the real-time empty of high resolution digital image Between territory filtering.

Background technology

High resolution digital image can provide abundant scene detailed information, thus becomes current medical image, remote sensing The urgent needs of the application such as remote measurement and multimedia communication.But digital picture is formed at it and holds very much during transmission Being vulnerable to the pollution of noise, the noisy digital picture of this band is unfavorable for being analyzed image, the most effectively filters and make an uproar Sound also retains one of the important subject that the information of original image is digital image processing field to greatest extent.

At these traditional filtering methods of gaussian filtering, medium filtering, mean filter and bilateral filtering, wherein Gauss filter Ripple, medium filtering and mean filter can effectively remove the noise in image, but these methods can not protect the edge of image Information, bilateral filtering method adds again brightness weights on the basis of considering distance weights, and bilateral filtering method can have Effect is removed noise and the marginal information of image has been carried out a certain degree of protection.Shimodaira H. is at IEEE " Patch-based has been delivered on International Conference on Image Processing.2013:868 871 Bilateral filter:Algorithms and performance " paper, literary composition proposes a kind of burst bilateral filtering Algorithm, compared with filtering method bilateral with tradition, when calculating codomain Gaussian template be by measure neighborhood of pixels block to be filtered with In search window, the structural similarity of other neighborhood of pixels block obtains brightness weights, and this method can effectively remove noise the most very well The marginal information protecting image.But, this algorithm computation complexity is higher, it is impossible to that applies at high-definition picture is real-time In process task.Additionally, existing burst bilateral filtering method can only operate in single architecture CPU, it is impossible on isomery framework Run, poor universality.

OpenCL (Open Computing Language) be first towards heterogeneous system general purpose multiple programming Standard, is also a unified programmed environment, it is adaptable to multi-core processor (CPU), graphic process unit (GPU) and scene can be compiled The parallel processors such as journey logical device (FPGA).Compared with tradition single architecture, isomery framework can preferably realize high-performance Parallel computation, in big data mining, the field such as machine learning and scientific algorithm holds out broad prospects.Image procossing comprise a large amount of The complex logic of the highly-parallel Floating-point Computation ability of GPU with CPU can be processed with the floating-point matrix computing of executed in parallel Combine with task scheduling ability and image procossing is accelerated.Professional platform independence good for OpenCL can also make program run On different platforms, portability and the versatility of code are enhanced.

Summary of the invention

It is an object of the invention to the defect overcoming above-mentioned prior art to exist, it is provided that a kind of GPU based on OpenCL Accelerate burst bilateral filtering method, by under OpenCL standard for GPU many PE unit calculating operation divide and scheduling, and Storage distribution optimizes, and completes the algorithm parallel acceleration under isomery framework, is used for solving to deposit in existing burst bilateral filtering method Filtration efficiency is low and the technical problem of poor universality.

For achieving the above object, the technical scheme that the present invention takes comprises the steps:

Step 1, in host side by universal cpu, selects the calculating equipment of heterogeneous computing platforms and its correspondence, at this isomery Calculate and on platform, create context, and create command queue in this context, obtain isomery framework.

Step 2, inputs original image, and host side reads this raw image data in host memory, simultaneously for processing after View data distribution host memory.

Step 3, host side utilizes Gaussian function to generate spatial domain Gaussian template, and this template is stored in host memory.

Step 4, host side creates multiple cache object in the context that step 1 obtains, and is respectively used to equipment end storage Image and spatial domain Gaussian template after original image, process, write original image and spatial domain Gaussian template by described command queue Enter the cache object of correspondence.

Step 5, utilizes OpenCL standard, and burst bilateral filtering algorithm is carried out parallel programming, by dividing after parallelization Sheet bilateral filtering algorithm is write as kernel function.

Step 6, sets the width of original image and height, spatial domain Gaussian template radius and multiple cache object in host side It is set to the parameter of kernel function.

Step 7, arranges the work item in the number of dimensions that kernel function falls in lines required, each dimension and each in host side The work item that working group processes.

Step 8, the parameter that host side is arranged according to step 7, kernel function of falling in lines.

Step 9, the spatial domain Gaussian template that equipment end generates according to host side determines search window size, in search window based on The brightness weights that structural similarity carries out calculating are as codomain Gaussian template.

Step 10, equipment end is by the data in cache object corresponding for original image, spatial domain Gaussian template and codomain Gauss Template is normalized weighted average calculation, the view data after being processed, and this view data writes the slow of its correspondence Deposit object.

Step 11, the cache object that after host side reading process, image is corresponding, the view data after being processed, and should View data writes host memory.

The present invention compared with prior art, has the advantage that

1, due to the fact that burst bilateral filtering method has been carried out GPU acceleration, the filter bilateral to burst with prior art Wave method calculates under single architecture CPU and compares, and GPU has carried out parallel computation to burst bilateral filtering method, and GPU is to floating-point The computing capability CPU the to be significantly larger than computing capability to floating-point, keeping, burst bilateral filtering method noise removal capability is constant In the case of, the calculating speed of burst bilateral filtering method is greatly improved.

2, due to the fact that have employed multiple programming standard based on OpenCL is carried out burst bilateral filtering method parallel Changing programming, compared with prior art, the present invention can realize burst bilateral filtering method and can transport under different isomerization platform OK, portability and the versatility of code are enhanced.

Accompanying drawing explanation

Fig. 1 is the FB(flow block) of the present invention；

Fig. 2 is the filter effect figure of the present invention and prior art.

Detailed description of the invention

Below in conjunction with the drawings and specific embodiments, the invention will be further described.

With reference to Fig. 1, the present invention comprises the following steps:

Step 1, calculates the calculating equipment of platform and its correspondence, can use the AMD of AMD isomery development platform and its correspondence GPU or NVIDIA isomery development platform and NVIDIA GPU or the AMD isomery development platform of correspondence and the CPU of correspondence, due to The calculating equipment that experiment main frame is installed is AMD R9 270X GPU, and the present embodiment uses AMD isomery development platform and its correspondence AMD R9 270X GPU, creates context and command queue in AMD isomery development platform, obtains isomery framework.

Step 2, chooses the gray level image of 8 as original image, and host side reads this raw image data in main frame In depositing, simultaneously for 8 greyscale image data distribution host memories after processing.

Step 3, host side utilizes Gaussian function to generate spatial domain Gaussian template, spatial domain Gaussian template and concrete pixel value without Closing, the most relevant with the relative position of pixel, for the calculating of all pixels, spatial domain Gaussian template is all fixing, so Host side generates spatial domain Gaussian template, and Gauss formula is:

$C_g(x,y)=e^{\{-\frac{{(x-x_0)}^2+{(y-y_0)}^2)}{2{\mathrm{\σ}}_d^2}\}}---\left(1\right)$

Wherein, (x₀,y₀) it is the coordinate of image slices vegetarian refreshments, (x, y) is the coordinate that spatial domain Gaussian template is corresponding,For spatial domain The variance of Gaussian template.

Arranging spatial domain Gaussian template size in the present embodiment is 7 × 7, the variance of spatial domain Gaussian templateIt is 10, will obtain Gaussian template be stored in host memory.

Step 4, host side creates multiple cache object within a context, is respectively used to equipment end and stores original 8 gray scales Image, process rear 8 gray level images and spatial domain Gaussian template, the caching that original 8 gray level images are corresponding with spatial domain Gaussian template Object Creation is read-only cache object, and the cache object processing rear 8 gray level images corresponding is created as only writing cache object, passes through Described command queue is by cache object corresponding with the write of spatial domain Gaussian template for original 8 gray level images.

Step 5, utilizes OpenCL standard, and burst bilateral filtering algorithm is carried out parallel programming, by dividing after parallelization Sheet bilateral filtering algorithm is write as kernel function.

Step 6, in host side by the width of original 8 gray level images and height, spatial domain Gaussian template radius and multiple slow Depositing object and be set to the parameter of kernel function, the caching that the most original 8 gray level images are corresponding with processing rear 8 gray level images is right As parameter is arranged in the global memory of equipment, the cache object parameter that spatial domain Gaussian template is corresponding is arranged in the constant of equipment In depositing, the width of original 8 gray level images be arranged on height, parameter that spatial domain Gaussian template radius is corresponding equipment privately owned in In depositing.

Step 7, sets kernel function required number of dimensions of falling in lines as 2 in host side, and the work item in each dimension is divided Not Wei pixel count on the width of original 8 gray level images and height, the work item that each working group processes is set to 16 × 16。

Step 8, host side is fallen in lines kernel function by command queue.

Step 9, the search window size that the spatial domain Gaussian template that equipment end generates according to host side determines is 7 × 7, search window In the brightness weights carrying out calculating based on structural similarity as codomain Gaussian template, its formula is:

$C_s(x,y)=e^{\{-\underset{i=-p}{\overset{p}{\Σ}}\underset{j=-p}{\overset{p}{\Σ}}\{\frac{{(I(x_0+i,y_0+j)-I(x+i,y+j))}^2}{2{\mathrm{\σ}}_r^2}\left\}\right\}}---\left(2\right)$

Wherein, (x y) is (x, y) gray value at place, (2p+1) × (2p+1) the burst neighbour described in image pixel point coordinates to I Territory size, (i, j) be in sheet pixel relative to sheet central point (x, coordinate offset y).

The size arranging burst neighborhood (2p+1) × (2p+1) in the present embodiment is 3 × 3, to (x, 3 y) in search window × 3 neighborhood sheets and center pixel (x₀,y₀) 3 × 3 neighborhood sheets do structural similarity calculate, using result of calculation as codomain Gauss Template.

Step 10, equipment end is by the data in cache object corresponding for original image, spatial domain Gaussian template and codomain Gauss Template is normalized weighted average calculation, and its formula is:

$J(x_0,y_0)=\frac{1}{K(x_0,y_0)}\underset{(x,y)\∈w(x_0,y_0)}{\Σ}{C}_g(x,y)C_s(x,y)I(x,y)---\left(3\right)$

$K(x_0,y_0)=\underset{(x,y)\∈w(x_0,y_0)}{\Σ}{C}_g(x,y)C_s(x,y)---\left(4\right)$

Wherein J (x₀,y₀) be burst bilateral filtering process after coordinate be (x₀,y₀) pixel value, w (x₀,y₀) for searching for Window, K (x₀,y₀) it is normalization factor.

The calculating of one pixel is carried out under a work item, and multiple work item perform to realize parallel computation simultaneously, obtain View data J (x after process₀,y₀), and this view data is write the cache object of its correspondence.

Step 11, the cache object that after host side reading process, image is corresponding, the view data after being processed, and should View data writes host memory.

Below in conjunction with the drawings and specific embodiments, the technique effect of the present invention is described in further detail:

1. experimental situation:

Heterogeneous platform is AMD isomery development platform, and wherein host side universal cpu is Intel Xeon E5410 CPU, meter Calculation equipment GPU is AMD R9 270X GPU, and software environment is Visual Studio 2013 and AMD APP SDK v2.9, behaviour It is Windows 7 as system.

2. experiment content:

8 gray noise images of high-resolution of input are used tradition burst bilateral filtering method and the inventive method respectively Being filtered operation, evaluate different images noise level and filtering speed respectively, its result is respectively such as Fig. 2 and Biao 1 institute Show.

Be raw noise image with reference to Fig. 2, Fig. 2 (a), its PSNR be 20.3643dB, Fig. 2 (b) be traditional bilateral filter of burst Design sketch after wave method denoising, its PSNR be 27.0669dB, Fig. 2 (c) be the design sketch after the inventive method denoising, its PSNR is 27.0669dB.

As seen from Figure 2, the inventive method is identical with tradition burst bilateral filtering method denoising performance.

Table 1 gives the calculating time of the inventive method and tradition burst bilateral filtering method, and burst size is 3 × 3, Search window size is 5 × 5 and 7 × 7.

The calculating time (unit: ms) of the different burst bilateral filtering method of table 1

By data in table it can be seen that GPU based on OpenCL proposed by the invention accelerates burst bilateral filtering method The real-time process to high-definition picture can be realized, and relatively conventional burst bilateral filtering method obtains the acceleration of highly significant Effect.

Claims (6)

1. a GPU based on OpenCL accelerates burst bilateral filtering method, it is characterised in that comprise the following steps:

(1), in host side by universal cpu, select the calculating equipment of heterogeneous computing platforms and its correspondence, put down in this Heterogeneous Computing Create context on platform, and create command queue in this context, obtain isomery framework；

(2), input original image, host side reads this raw image data in host memory, simultaneously for the image after processing Data distribution host memory；

(3), host side utilizes Gaussian function to generate spatial domain Gaussian template, and this template is stored in host memory；

(4), host side in the context that step (1) obtains, create multiple cache object, be respectively used to equipment end storage original Image and spatial domain Gaussian template after image, process, it is right original image and spatial domain Gaussian template to be write by described command queue The cache object answered；

(5), utilize OpenCL standard, burst bilateral filtering algorithm is carried out parallel programming, by bilateral for the burst after parallelization Filtering algorithm is write as kernel function；

(6), in the width of original image and height, spatial domain Gaussian template radius and multiple cache object are set to by host side The parameter of kernel function；

(7), the work item in the number of dimensions that kernel function falls in lines required, each dimension and each working group are set in host side The work item processed；

(8), the parameter that arranges according to step (7) of host side, kernel function of falling in lines；

(9), the spatial domain Gaussian template that generates according to host side of equipment end determine search window size, in search window based on structure phase Seemingly spend the brightness weights carrying out calculating as codomain Gaussian template；

(10), the data in cache object corresponding for original image, spatial domain Gaussian template and codomain Gaussian template are entered by equipment end Row normalization weighted average calculation, the view data after being processed, and this view data is write the cache object of its correspondence；

(11), image is corresponding after host side reading process cache object, the view data after being processed, and by this picture number According to writing host memory.

GPU based on OpenCL the most according to claim 1 accelerates burst bilateral filtering method, it is characterised in that: step (3) Gaussian function described in, it is expressed as:

$C_g(x,y)=e^{\{-\frac{{(x-x_0)}^2+{(y-y_0)}^2)}{2{\mathrm{\σ}}_d^2}\}}$

Wherein, (x₀,y₀) it is the coordinate of image slices vegetarian refreshments, (x, y) is the coordinate that spatial domain Gaussian template is corresponding,For spatial domain Gauss The variance of template.

GPU based on OpenCL the most according to claim 1 accelerates burst bilateral filtering method, it is characterised in that: step (4) the multiple cache objects described in, the cache object that wherein original image is corresponding with spatial domain Gaussian template is all set as read-only Cache object, after process, image correspondence cache object is set as only writing cache object.

GPU based on OpenCL the most according to claim 1 accelerates burst bilateral filtering method, it is characterised in that: step (6) parameter of the kernel function described in, the cache object parameter that wherein original image is corresponding with image after process is arranged on and sets In standby global memory, the cache object parameter that spatial domain Gaussian template is corresponding is arranged in the constant internal memory of equipment, original image Width be arranged in the privately owned internal memory of equipment with height, parameter that spatial domain Gaussian template radius is corresponding.

GPU based on OpenCL the most according to claim 1 accelerates burst bilateral filtering method, it is characterised in that: step (9) calculating based on structural similarity described in, its formula is:

$C_s(x,y)=e^{\{-\underset{i=-p}{\overset{p}{\mathrm{\Σ}}}\underset{j=-p}{\overset{p}{\mathrm{\Σ}}}\{\frac{{\left(I\right(x_0+i,y_0+j)-I(x+i,y+j\left)\right)}^2}{2{\mathrm{\σ}}_r^2}\left\}\right\}}$

Wherein, I (x, y) be described in image pixel point coordinates (x, y) gray value at place, (2p+1) × (2p+1) burst neighborhood is big Little, (i, j) be in sheet pixel relative to sheet central point (x, coordinate offset y).

GPU based on OpenCL the most according to claim 1 accelerates burst bilateral filtering method, it is characterised in that: step (10) the normalization weighted average calculation described in, its formula is:

$J(x_0,y_0)=\frac{1}{K(x_0,y_0)}\underset{(x,y)\∈w(x_0,y_0)}{\mathrm{\Σ}}{C}_g(x,y)C_s(x,y)I(x,y)$

$K(x_0,y_0)=\underset{(x,y)\∈w(x_0,y_0)}{\mathrm{\Σ}}{C}_g(x,y)C_s(x,y)$

Wherein J (x₀,y₀) be burst bilateral filtering process after coordinate be (x₀,y₀) pixel value, w (x₀,y₀) it is search window, K (x₀,y₀) it is normalization factor.